Noise is one of the basic factors that degrade remote sensing (RS) data and prevent accurate and reliable retrieval of useful information. Availability of a priori information on noise type and properties allows applying more effective methods for image processing, namely, filtering, edge detection, feature extraction, etc. However, noise statistics are often unknown and are to be estimated for an image at hand. Thus one needs blind methods for the evaluation of the noise variance, especially if the number of images or sub-band images of multichannel RS data is large enough. In this paper, we consider several approaches to blind evaluation of noise variance. An important item is that we consider both i.i.d. and spatially correlated noise. It is demonstrated that some techniques that perform well enough for i.i.d. noise fail if the image is corrupted by spatially correlated noise. We show how segmentation-based methods for blind evaluation of noise variance that operate in the spatial domain can be modified in order to provide better accuracy for wide ranges of noise variance and spatial correlation parameters. Numerical simulation results comparing the performance of several techniques are presented. Real RS data processing examples are also given.